build_mixed_dataset.py

#!/usr/bin/env python3
"""
build_mixed_dataset.py — Four-source mixed pre-training corpus builder.

Sources
-------
  SLTrans      local parquet  (balanced language x IR-type)   --sltrans-tokens
  peS2o        allenai/peS2o  (open scientific papers)        --pes2o-tokens
  TheStack     bigcode/the-stack (permissively licensed code)  --stack-tokens
  OpenWebMath  open-web-math/open-web-math (math web text)    --owm-tokens

Set any cap to 0 to skip that source.

Output
------
  JSONL shards under --output-dir, one filename prefix per source:
    sltrans-00000.jsonl, pes2o-00000.jsonl, the_stack-00000.jsonl, ...
  Each record:
    {"text": "...", "source": "...", "meta": {...}, "est_tokens": N}
  Plus manifest.json summarising the run.

Usage
-----
  pip install "datasets>=2.18" pyarrow pandas tqdm
  huggingface-cli login          # peS2o and the-stack are gated

  python build_mixed_dataset.py
  python build_mixed_dataset.py --sltrans-tokens 500e6 --owm-tokens 200e6
  python build_mixed_dataset.py --stack-tokens 0      # skip TheStack
  python build_mixed_dataset.py --stack-langs python,rust,go
"""

from __future__ import annotations

import argparse
import json
import random
import re
import socket
import sys
import time
from concurrent.futures import ThreadPoolExecutor
from pathlib import Path

import pandas as pd
import pyarrow.parquet as pq
from tqdm import tqdm

socket.setdefaulttimeout(90)

# ── constants ──────────────────────────────────────────────────────────────────
SLTRANS_PROBE_ROWS = 200
SLTRANS_SKIP_DIRS  = {".venv", "__pycache__", ".git"}

THE_STACK_LANGS = [
    "python", "c", "cpp", "rust", "go",
    "java", "javascript", "typescript",
]

_TRANSIENT_ERRORS = ("ssl", "timeout", "handshake", "connection", "timed out")


# ── token estimation ───────────────────────────────────────────────────────────

def estimate_tokens(text: str) -> int:
    return int(len(text.split()) * 1.5)


# ── JSONL shard writer ─────────────────────────────────────────────────────────

class ShardWriter:
    def __init__(self, out_dir: Path, prefix: str, records_per_shard: int):
        out_dir.mkdir(parents=True, exist_ok=True)
        self._dir, self._pfx, self._rps = out_dir, prefix, records_per_shard
        self._idx = self._n = 0
        self._fh = None
        self._roll()

    def _roll(self):
        if self._fh:
            self._fh.close()
        self._fh = (self._dir / f"{self._pfx}-{self._idx:05d}.jsonl").open("w", encoding="utf-8")
        self._n = 0
        self._idx += 1

    def write(self, record: dict):
        self._fh.write(json.dumps(record, ensure_ascii=False) + "\n")
        self._n += 1
        if self._n >= self._rps:
            self._roll()

    def close(self):
        if self._fh:
            self._fh.close()
            self._fh = None


# ── SLTrans (local parquet) ────────────────────────────────────────────────────

def _sltrans_find_groups(root: Path) -> dict[tuple[str, str], list[Path]]:
    """Return {(language, ir_type): [sorted shard paths]}."""
    groups: dict[tuple[str, str], list[Path]] = {}
    for d in sorted(root.iterdir()):
        if not d.is_dir() or d.name in SLTRANS_SKIP_DIRS:
            continue
        for f in sorted(d.glob("*.parquet")):
            m = re.match(r"^(Perf_Optimized|Size_Optimized)", f.name)
            if m:
                groups.setdefault((d.name, m.group(1)), []).append(f)
    return groups


def _pq_nrows(files: list[Path]) -> int:
    return sum(pq.ParquetFile(f).metadata.num_rows for f in files)


def _est_tok_df(src: pd.Series, ir: pd.Series) -> pd.Series:
    src_w = src.fillna("").str.split().str.len().fillna(0)
    ir_w  = ir.fillna("").str.split().str.len().fillna(0)
    return ((src_w + ir_w + 5) * 1.5).astype(int)


def _probe_avg_tokens(files: list[Path], n: int, rng: random.Random) -> float:
    frames = []
    seed = rng.randint(0, 2**31)
    for f in files:
        df = pq.ParquetFile(f).read_row_group(0).to_pandas()
        if not df.empty:
            frames.append(df.sample(min(n, len(df)), random_state=seed))
        if sum(len(x) for x in frames) >= n:
            break
    if not frames:
        return 0.0
    p = pd.concat(frames, ignore_index=True).head(n)
    p = p.dropna(subset=["Source_Code", "IR_Original"])
    p = p[(p["Source_Code"] != "") & (p["IR_Original"] != "")]
    return float(_est_tok_df(p["Source_Code"], p["IR_Original"]).mean()) if len(p) else 0.0


def _sltrans_allocate(
    groups: dict[tuple[str, str], list[Path]],
    total: int,
    rng: random.Random,
) -> dict[tuple[str, str], int]:
    """Equal-share budget with deficit redistribution for small groups."""
    keys = sorted(groups)
    avail: dict[tuple[str, str], int] = {}
    for k in tqdm(keys, desc="      probe", unit="grp", leave=False):
        rows = _pq_nrows(groups[k])
        avg  = _probe_avg_tokens(groups[k], SLTRANS_PROBE_ROWS, rng)
        avail[k] = int(rows * avg)
        tqdm.write(
            f"      {k[0]:>15}/{k[1]:<16}  ~{avail[k]:>14,} tok"
            f"  ({rows:,} rows, avg {avg:.0f})"
        )
    budgets = {k: total // len(keys) for k in keys}
    for _ in range(len(keys)):
        capped  = {k: min(budgets[k], avail[k]) for k in keys}
        deficit = sum(budgets[k] - capped[k] for k in keys)
        if not deficit:
            break
        room = [k for k in keys if capped[k] < avail[k]]
        if not room:
            break
        bonus = deficit // len(room)
        for k in room:
            capped[k] = min(capped[k] + bonus, avail[k])
        budgets = capped
    return budgets


def write_sltrans(
    root: Path,
    budget: int,
    writer: ShardWriter,
    rng: random.Random,
    min_tokens: int,
) -> int:
    groups = _sltrans_find_groups(root)
    if not groups:
        print(f"    WARNING: no SLTrans parquet files found in {root}", file=sys.stderr)
        return 0

    budgets = _sltrans_allocate(groups, budget, rng)
    total_written = 0
    bar = tqdm(total=budget, unit="tok", unit_scale=True,
               desc="      write", dynamic_ncols=True)

    for (lang, ir_type) in sorted(groups):
        g_budget  = budgets[(lang, ir_type)]
        g_written = 0
        files = list(groups[(lang, ir_type)])
        rng.shuffle(files)

        for f in files:
            if g_written >= g_budget:
                break
            pf = pq.ParquetFile(f)
            for gi in range(pf.num_row_groups):
                if g_written >= g_budget:
                    break
                df = pf.read_row_group(gi).to_pandas()
                df = df.dropna(subset=["Source_Code", "IR_Original"])
                df = df[(df["Source_Code"] != "") & (df["IR_Original"] != "")]
                if df.empty:
                    continue
                df = df.sample(frac=1, random_state=rng.randint(0, 2**31)).reset_index(drop=True)
                df["_t"] = _est_tok_df(df["Source_Code"], df["IR_Original"])

                remaining = g_budget - g_written
                cutoff = max(int((df["_t"].cumsum() <= remaining).sum()), 1)
                for row in df.iloc[:cutoff].to_dict("records"):
                    toks = int(row["_t"])
                    if toks < min_tokens:
                        continue
                    text = (
                        f"<source>\n{row['Source_Code']}\n</source>\n"
                        f"<llvm_ir>\n{row['IR_Original']}\n</llvm_ir>"
                    )
                    writer.write({
                        "text": text,
                        "source": "sltrans",
                        "meta": {"language": lang, "ir_type": ir_type},
                        "est_tokens": toks,
                    })
                    g_written    += toks
                    total_written += toks
                    bar.update(min(toks, budget - bar.n))
    bar.close()
    return total_written


# ── HuggingFace streaming ──────────────────────────────────────────────────────

def _hf_open(
    hf_path: str,
    split: str = "train",
    hf_config: str | None = None,
    data_dir: str | None = None,
):
    """Open one HF streaming dataset with exponential-backoff retry."""
    from datasets import load_dataset

    kw: dict = {"split": split, "streaming": True}
    if hf_config:
        kw["name"] = hf_config
    if data_dir:
        kw["data_dir"] = data_dir

    for attempt in range(5):
        try:
            return load_dataset(hf_path, **kw)
        except ValueError as e:
            if "Bad split" in str(e):
                return None
            raise
        except Exception as e:
            if attempt < 4 and any(k in str(e).lower() for k in _TRANSIENT_ERRORS):
                time.sleep(2 ** attempt)
                continue
            raise
    return None


def _hf_iter(
    hf_path: str,
    split: str = "train",
    hf_config: str | None = None,
    subsets: list[str] | None = None,
):
    """
    Yield rows from a HuggingFace streaming dataset.
    For TheStack, pass subsets; streams are resolved in parallel and interleaved.
    """
    if not subsets:
        ds = _hf_open(hf_path, split=split, hf_config=hf_config)
        if ds is not None:
            yield from ds
        return

    # Resolve subset streams in parallel (each resolution is an HTTP round-trip).
    def _open_sub(sub: str):
        return _hf_open(hf_path, split=split, data_dir=f"data/{sub}")

    with ThreadPoolExecutor(max_workers=min(4, len(subsets))) as pool:
        streams = [s for s in pool.map(_open_sub, subsets) if s is not None]

    if streams:
        from datasets import interleave_datasets
        yield from interleave_datasets(streams, stopping_strategy="all_exhausted")


def write_hf_source(
    source_name: str,
    budget: int,
    writer: ShardWriter,
    rng: random.Random,
    min_tokens: int,
    hf_path: str,
    text_fn,
    meta_fn,
    hf_config: str | None = None,
    split: str = "train",
    subsets: list[str] | None = None,
) -> int:
    written = skipped = 0
    bar = tqdm(total=budget, unit="tok", unit_scale=True,
               desc=f"      {source_name:<12}", dynamic_ncols=True, smoothing=0.05)
    try:
        for row in _hf_iter(hf_path, split=split, hf_config=hf_config, subsets=subsets):
            text = text_fn(row)
            if not text:
                skipped += 1
                continue
            toks = estimate_tokens(text)
            if toks < min_tokens:
                skipped += 1
                continue
            writer.write({
                "text": text,
                "source": source_name,
                "meta": meta_fn(row),
                "est_tokens": toks,
            })
            written += toks
            bar.update(min(toks, budget - bar.n))
            if written >= budget:
                break
    finally:
        bar.close()
    print(f"      done: {written:,} tokens written, {skipped:,} rows skipped")
    return written


# ── text / meta extractors ─────────────────────────────────────────────────────

def _get(row: dict, *keys: str, default: str = "") -> str:
    for k in keys:
        v = row.get(k)
        if v:
            return str(v)
    return default


def pes2o_text(row):  return _get(row, "text", "content")
def pes2o_meta(row):  return {"id": _get(row, "id", "doc_id"), "source": _get(row, "source", "venue")}

def stack_text(row):  return _get(row, "content", "text", "code")
def stack_meta(row):  return {
    "lang":    _get(row, "lang", "language"),
    "repo":    _get(row, "max_stars_repo_name", "repo_name"),
    "license": _get(row, "license"),
}

def owm_text(row):    return _get(row, "text")
def owm_meta(row):    return {"url": _get(row, "url")}


# ── main ───────────────────────────────────────────────────────────────────────

def main() -> None:
    ap = argparse.ArgumentParser(
        description=__doc__,
        formatter_class=argparse.RawDescriptionHelpFormatter,
    )
    ap.add_argument("--sltrans-root",   default=".",
                    help="Root dir of downloaded SLTrans parquet files (default: .)")
    ap.add_argument("--sltrans-tokens", type=float, default=700_000_000,
                    help="Token cap for SLTrans   (default: 700M, 0=skip)")
    ap.add_argument("--pes2o-tokens",   type=float, default=150_000_000,
                    help="Token cap for peS2o     (default: 150M, 0=skip)")
    ap.add_argument("--stack-tokens",   type=float, default=100_000_000,
                    help="Token cap for TheStack  (default: 100M, 0=skip)")
    ap.add_argument("--owm-tokens",     type=float, default=50_000_000,
                    help="Token cap for OpenWebMath (default: 50M, 0=skip)")
    ap.add_argument("--stack-langs",    default=",".join(THE_STACK_LANGS),
                    help="Comma-separated TheStack language subsets")
    ap.add_argument("--output-dir",     default="./mixed_pretrain",
                    help="Output directory for JSONL shards (default: ./mixed_pretrain)")
    ap.add_argument("--shard-size",     type=int, default=50_000,
                    help="Records per JSONL shard (default: 50000)")
    ap.add_argument("--min-tokens",     type=int, default=32,
                    help="Drop records shorter than this (est. tokens, default: 32)")
    ap.add_argument("--seed",           type=int, default=42)
    args = ap.parse_args()

    rng         = random.Random(args.seed)
    out_dir     = Path(args.output_dir)
    stack_langs = [s.strip() for s in args.stack_langs.split(",") if s.strip()]

    budgets = {
        "sltrans":     int(args.sltrans_tokens),
        "pes2o":       int(args.pes2o_tokens),
        "the_stack":   int(args.stack_tokens),
        "openwebmath": int(args.owm_tokens),
    }
    active_sources = [name for name, tok in budgets.items() if tok > 0]
    total_budget   = sum(budgets.values())

    print("=" * 64)
    print("Mixed pre-training dataset builder")
    print(f"  Output : {out_dir.resolve()}")
    print(f"  Seed   : {args.seed}")
    print()
    for name, toks in budgets.items():
        if toks > 0:
            print(f"  {name:<14}  {toks:>15,} tokens")
        else:
            print(f"  {name:<14}  (skipped)")
    print(f"  {'TOTAL':<14}  {total_budget:>15,} tokens")
    print("=" * 64)

    summary: dict[str, int] = {}
    n_active = len(active_sources)
    step = 1

    # ── SLTrans ────────────────────────────────────────────────────────────────
    if budgets["sltrans"] > 0:
        print(f"\n[{step}/{n_active}] SLTrans  (local parquet, balanced language x IR-type)")
        step += 1
        w = ShardWriter(out_dir, "sltrans", args.shard_size)
        try:
            summary["sltrans"] = write_sltrans(
                Path(args.sltrans_root), budgets["sltrans"], w, rng, args.min_tokens,
            )
        finally:
            w.close()

    # ── peS2o ──────────────────────────────────────────────────────────────────
    if budgets["pes2o"] > 0:
        print(f"\n[{step}/{n_active}] peS2o  (allenai/peS2o, config=v2)")
        step += 1
        w = ShardWriter(out_dir, "pes2o", args.shard_size)
        try:
            summary["pes2o"] = write_hf_source(
                "pes2o", budgets["pes2o"], w, rng, args.min_tokens,
                hf_path="allenai/peS2o",
                text_fn=pes2o_text, meta_fn=pes2o_meta,
                hf_config="v2",
            )
        finally:
            w.close()

    # ── TheStack ───────────────────────────────────────────────────────────────
    if budgets["the_stack"] > 0:
        print(f"\n[{step}/{n_active}] TheStack  (bigcode/the-stack, {len(stack_langs)} language subsets)")
        print(f"      langs: {', '.join(stack_langs)}")
        step += 1
        w = ShardWriter(out_dir, "the_stack", args.shard_size)
        try:
            summary["the_stack"] = write_hf_source(
                "the_stack", budgets["the_stack"], w, rng, args.min_tokens,
                hf_path="bigcode/the-stack",
                text_fn=stack_text, meta_fn=stack_meta,
                subsets=stack_langs,
            )
        finally:
            w.close()

    # ── OpenWebMath ────────────────────────────────────────────────────────────
    if budgets["openwebmath"] > 0:
        print(f"\n[{step}/{n_active}] OpenWebMath  (open-web-math/open-web-math)")
        w = ShardWriter(out_dir, "openwebmath", args.shard_size)
        try:
            summary["openwebmath"] = write_hf_source(
                "openwebmath", budgets["openwebmath"], w, rng, args.min_tokens,
                hf_path="open-web-math/open-web-math",
                text_fn=owm_text, meta_fn=owm_meta,
            )
        finally:
            w.close()

    # ── manifest ───────────────────────────────────────────────────────────────
    manifest = {
        "seed":                    args.seed,
        "min_tokens_per_record":   args.min_tokens,
        "sources": {
            "sltrans":     {"root":     args.sltrans_root,                 "target_tokens": budgets["sltrans"]},
            "pes2o":       {"hf_path":  "allenai/peS2o",                  "target_tokens": budgets["pes2o"]},
            "the_stack":   {"hf_path":  "bigcode/the-stack",              "target_tokens": budgets["the_stack"],   "langs": stack_langs},
            "openwebmath": {"hf_path":  "open-web-math/open-web-math",    "target_tokens": budgets["openwebmath"]},
        },
        "tokens_written": summary,
    }
    (out_dir / "manifest.json").write_text(json.dumps(manifest, indent=2))

    # ── summary ────────────────────────────────────────────────────────────────
    grand = sum(summary.values())
    print("\n" + "=" * 62)
    print(f"{'Source':<14}  {'Target':>15}  {'Written':>15}  {'Share':>6}")
    print("-" * 58)
    for name in ["sltrans", "pes2o", "the_stack", "openwebmath"]:
        if budgets[name] == 0:
            continue
        written = summary.get(name, 0)
        pct     = 100 * written / grand if grand else 0
        print(f"{name:<14}  {budgets[name]:>15,}  {written:>15,}  {pct:>5.1f}%")
    print("-" * 58)
    print(f"{'TOTAL':<14}  {total_budget:>15,}  {grand:>15,}  100.0%")
    print(f"\nOutput: {out_dir.resolve()}")


if __name__ == "__main__":
    main()